Implementation checks
These records check the implementation, not the model. They show that the code reproduces the published calculations and recovers known parameters from synthetic data; they do not validate the alpha law, HICUP or the dual-node model as FinFET rate models, which has not yet been done (see validate and plan the next runs on the tool page).
These are the check records the methodology refers to, published as the harnesses write them so that every claim of agreement on the tool page can be read against its number. Each record is regenerated by the script named with it whenever the engine changes; PASS and FAIL lines are the script's own. The harnesses, like the engines, were written with Claude, Anthropic's AI model, and reviewed by hand (how this site is made); a harness written by the same hand as the code catches transcription errors against the paper's numbers, not misreadings shared by both, which is why a reader's check is asked for. The data behind them are public: Edmonds' 2002 paper, the 1995 HICUP paper, JPL Publication 11-6 and the Versal worked example on the tool page. Anyone who finds a check that does not hold, or a check that is missing, is asked to say so through the community forum.
The site's rate integrator against a line for line port of the DIAL_E FORTRAN listing in Edmonds' 2002 paper, on the same LET table and inputs.
dial_e_port.py is a line for line Python port of the FORTRAN listing in Appendix C of Edmonds, IEEE TNS 49(3), 1522, 2002.
Run: python3 dial_e_port.py env_fx.json GCR_SolMin 0.852 17.04 0 (sigma_sat cm2, L1/e, cutoff LET; the paper's Toshiba DRAM values).
Same inputs, site engine (assets/alphalaw.js, AL.rate with the exponential form) on the same 210 point LET table:
| A, B | DIAL_E port | site engine, table grid | 400 pt grid | 2000 pt grid | converged |
|---|---|---|---|---|---|
| 2, 0.548 | 5.2092e-1 /d | 5.2092e-1 | 5.1350e-1 | 5.024e-1 | 4.9932e-1 |
| 1, 0 | 1.0852e+0 | 1.0852e+0 | 1.0711e+0 | ||
| 0, 0 cosine | 2.5637e+0 | 2.5637e+0 | 2.5361e+0 | ||
| 1, 1 iso | 6.2449e-1 | 6.2449e-1 | 6.1576e-1 |
Identical to five digits on the same grid. The summation by parts is conservative on coarse grids as Edmonds notes (6 percent high on the 210 point table, 3 percent on 400 points, under 0.5 percent on 2000, which the page now uses). Table I of the paper (0.0667 and 0.189 per device day) is a solar maximum spectrum, Z 2 to 92, not bundled here, so it was not reproduced numerically; the ratio of the two cases differs between spectra.
Defects in the printed listing: none in the arithmetic. Typesetting only: STATUS='OLD,' FILE='DIAL.IN' and
'RATE=,'RATE,'/DAY' have the comma moved inside the closing quote (American punctuation applied to code), and the
quotes are typographic. EXP(-1.0*L1OE/ARG) is correct in the PDF; a text extraction renders the minus as 01.0.
Changes made to the tool from this review: cutoff LET input (DIAL_E's LCUT with its defaults), boundary term H(L1) sigma_avg(L1) kept because the site tables start at 0.1 not 0.001 MeV cm2/mg, 2000 point grid, coverage note (L over alpha range the data constrain, share of the rate from below it), a DIAL_T table curve option (isotonic log log table through the collapsed points, A and B from the collapse scan refined by simplex, nonparametric bootstrap), and the methodology text on the Weibull as a fitting function and on the example's curve. Versal example, GCR solar minimum: fitted Weibull 2.79e-11 per bit day, table curve 3.16e-11 (ratio 0.88).
Attribution note (2026-09-08): A and B above one is not a Space RHA addition. Patterson and Edmonds (RADECS 2002) scanned A and B from 0 to 10 and reported a 512 Mbit SDRAM at A = 1.38, B = 6.56, noting that effective LET then falls with tilt. The page now credits them; what this site adds is the application to fins. External sources now cited for the other additions: Efron 1979 and Efron and Tibshirani 1993 (bootstrap); Ladbury 2007 and 2022, Hansen 2020, Sukhaseum et al. 2011 (Monte Carlo and Poisson resampling of SEE cross section data and rate uncertainty); Petersen 2008 (the Weibull as the field's convention); Ayer et al. 1955 and Barlow et al. 1972 (pool adjacent violators); Nelder and Mead 1965 (simplex). The interpolated LET grid has no source beyond Edmonds' own remark on grid spacing.
The transcription of the 1995 HICUP model (hicup/hicup.js): geometry identities, a Monte Carlo of chords, the incomplete gamma average, the reductions and limits, the paper's Table I device against its Figure 3, the angle average, the rate, and parameter recovery from synthetic counts. Regenerated by hicup/test_hicup.js.
Parameter sets: HC6856 (Figure 2 caption: L0 = 12.47, WL = 47.0), RK1-05 and 6504RH (Table I), a cube and an oblong cell.
PASS sigma(theta = 0) equals A F((L - L0) / WL) for five parameter sets and eight LETs: largest relative difference 4.24e-15
PASS (a) Ap - g1 h s + g2 s^2 equals the projected area of the box shrunk by s along the beam (10000 random cases): largest relative difference 2.11e-13 PASS (b) Am from the shrunken box equals the piecewise (A7): largest relative difference 1.79e-13 PASS (c) theta_c of (A6) is the switch of the longest chord (2000 cases): 0 mismatches PASS (d) Petroff's C0, C1 equal g1 h / Ap and 2 g2 / Ap: largest relative difference 2.07e-8 (the printed form cancels 1 - 2 Sbar / Smx + Am / Ap when the tilt is small) PASS (d) Petroff's distribution has unit weight and mean Sbar = l w h / Ap (A8): largest deviation 9.66e-16
PASS HC6856 tilt 30 rot 20: P(chord >= Sc) from 400000 chords matches (Ap - G) / Ap within 4 sigma: Sc/Smx 0.15: MC 0.9714 model 0.9714; Sc/Smx 0.4: MC 0.9243 model 0.9243; Sc/Smx 0.7: MC 0.8687 model 0.8687; Sc/Smx 0.95: MC 0.8229 model 0.8232; worst 0.37 sigma PASS HC6856 tilt 30 rot 20: mean chord 0.7325 um against Sbar 0.7324: Smx 0.808 um, delta weight Am / Ap 0.814 PASS HC6856 tilt 70 rot 45: P(chord >= Sc) from 400000 chords matches (Ap - G) / Ap within 4 sigma: Sc/Smx 0.15: MC 0.8831 model 0.8832; Sc/Smx 0.4: MC 0.7028 model 0.7028; Sc/Smx 0.7: MC 0.5104 model 0.5101; Sc/Smx 0.95: MC 0.3699 model 0.3693; worst 0.80 sigma PASS HC6856 tilt 70 rot 45: mean chord 1.3258 um against Sbar 1.3256: Smx 2.047 um, delta weight Am / Ap 0.343 PASS RK105 tilt 60 rot 0: P(chord >= Sc) from 400000 chords matches (Ap - G) / Ap within 4 sigma: Sc/Smx 0.15: MC 0.8657 model 0.8663; Sc/Smx 0.4: MC 0.6420 model 0.6434; Sc/Smx 0.7: MC 0.3766 model 0.3759; Sc/Smx 0.95: MC 0.1538 model 0.1530; worst 1.83 sigma PASS RK105 tilt 60 rot 0: mean chord 3.2132 um against Sbar 3.2145: Smx 5.800 um, delta weight Am / Ap 0.108 PASS CUBE tilt 50 rot 30: P(chord >= Sc) from 400000 chords matches (Ap - G) / Ap within 4 sigma: Sc/Smx 0.15: MC 0.7667 model 0.7667; Sc/Smx 0.4: MC 0.4455 model 0.4439; Sc/Smx 0.7: MC 0.1658 model 0.1652; Sc/Smx 0.95: MC 0.0238 model 0.0236; worst 2.00 sigma PASS CUBE tilt 50 rot 30: mean chord 1.8755 um against Sbar 1.8720: Smx 4.767 um, delta weight Am / Ap 0.005 PASS OBLONG tilt 75 rot 80: P(chord >= Sc) from 400000 chords matches (Ap - G) / Ap within 4 sigma: Sc/Smx 0.15: MC 0.8112 model 0.8112; Sc/Smx 0.4: MC 0.5388 model 0.5387; Sc/Smx 0.7: MC 0.2820 model 0.2815; Sc/Smx 0.95: MC 0.1254 model 0.1254; worst 0.64 sigma PASS OBLONG tilt 75 rot 80: mean chord 1.2357 um against Sbar 1.2354: Smx 2.575 um, delta weight Am / Ap 0.100
PASS gamma(a, x) for a = 1, 2, 3 against closed forms at seven x: largest relative difference 8.14e-13 PASS gamma(1/2, x) against sqrt(pi) erf(sqrt x) (erf to 1e-7): largest relative difference 9.01e-8 PASS series and continued fraction agree at the same x around the switch: largest relative difference 2.54e-15 PASS x beyond the double range (the sharp threshold limit) returns P = 1
PASS closed form equals quadrature for 150 cases (five sets, six directions, five LETs): largest relative difference 3.75e-5 (Simpson's own error at the kink of the toe for b = 1.3; the RK1-05 cases agree to 1e-10); example RK1-05 60 deg, L = 57.0: closed 1.21665e-7 quadrature 1.21665e-7
PASS b = 400: sigma tends to the point value at E = E0 + W (4 cases with the step inside the chord range; below it both vanish): largest relative difference 9.70e-4 PASS h / w = 0.001, Sf = 0: sigma equals the cosine law cos(theta) sigmaN(L / cos theta) to 0.3 percent: largest relative difference 1.54e-3 note: at h / w = 0.01 the side faces already move the 60 degree cross section 1.5 percent off the cosine law. note: with Sf = h the cosine law overstates the 60 degree cross section at L = 8 by a factor 2.42 (Leff of equation 16 is 12.0 against L / cos theta = 16) PASS scaled points (sigma_eff, Leff) fall on the normal incidence Weibull: largest relative difference 1.74e-15; the inverse cosine law leaves a scatter of up to 108 percent on the same points PASS at phi = 0 the correction is equation (14) with the bridging function (15): largest relative difference 3.05e-16
RK1-05 (Table I: h 2.9 um, sigma_sat 3.9e-7 cm2, L0 38, Sf 0.5 um, WL 43.4, b 2.04). One LET per ion chosen by least squares on the plotted curve: Xe 63.1, Kr 39.9, Cu 30.5 MeV cm2/mg. PASS Figure 3 curves reproduced from Table I: rms log residual 4 percent, worst 8 percent, over 11 readable points and two zeros: Xe 0 deg: model 1.09e-7 plot 1.10e-7; Xe 10 deg: model 1.08e-7 plot 1.10e-7; Xe 40 deg: model 1.42e-7 plot 1.35e-7; Xe 55 deg: model 1.56e-7 plot 1.55e-7; Xe 63 deg: model 1.56e-7 plot 1.58e-7; Kr 37 deg: model 1.20e-8 plot 1.20e-8; Kr 48 deg: model 2.66e-8 plot 2.80e-8; Kr 60 deg: model 4.97e-8 plot 4.60e-8; Kr 63 deg: model 5.52e-8 plot 5.30e-8; Cu 53 deg: model 5.14e-9 plot 5.00e-9; Cu 60 deg: model 1.23e-8 plot 1.30e-8; Kr 0 deg: model 6.95e-10 plot 0; Cu 45 deg: model 7.95e-10 plot 0 Table I is printed to three figures and the Kr curve at small tilt sits on the Weibull toe (L barely above L0), where the model changes 20 percent for a 2 percent change in the ion's LET, so the residuals are at the level of the rounding. The tilt rise of the Xe curve (1.1e-7 to 1.6e-7 by 60 degrees) is the side faces of a cell with h / w = 0.46 coming into the projected area; a model without the side faces would fall with tilt. HC6856 (Figure 2 caption). LET solved from the curves at 0 deg: Xe 69.3, Kr 44.2, Cu 34.0. Model at 45 and 60 deg: Xe 1.72e-7, 1.38e-7 (plot about 1.6e-7 and 1.15e-7); Kr 8.29e-8, 8.51e-8 (plot 0.72e-7 and 0.69e-7); Cu 3.91e-8, 4.64e-8 (plot 0.34e-7 and 0.36e-7). The Xe and Kr curves of Figure 2 sit below the transcription at 60 degrees by 15 to 20 percent while Figure 3, drawn from the same code with Table I, is reproduced. The Figure 2 data themselves (Xe 1.38e-7 at 60 deg, Kr 0.78e-7 at 60 deg) lie on the transcription. The caption's parameters are rounded and the ion LETs are not given, so the curves of Figure 2 cannot be pinned further; the equations, not the figure, are what is implemented.
PASS angle average saturates at Abar = (lw + lh + wh) / 2 (Figure 9): average at L = 2000: 1.6002e-7, Abar 1.6000e-7 cm2, normal incidence A 2.50e-7 PASS nothing upsets below Lmin of equation (21) and something does above: Lmin 2.62 MeV cm2/mg (L0 12.47), average at 1.3 Lmin 8.38e-14 Figure 9 check: the average reaches 74 percent of Abar at L = 60 and 100 percent at L = 140 (the plot shows the dotted curve near 1.5e-7 at 140). PASS rate by (19) equals the classic sum over thresholds (18) with 120 Weibull slices: GCR solar minimum: (19) 4.8590e-9, (18) 4.8663e-9 upsets per bit day PASS angle quadrature converged (30, 45, 90 segments within 0.5 percent): 30: 4.8559e-9, 45: 4.8590e-9, 90: 4.8609e-9 Against the SEE rate tool's mean chord effective flux (every chord in a direction replaced by the mean chord), square cell, Sf = 0, L0 5, WL 20, b 2, GCR solar minimum: Z/X 0.001: I-HICUP 1.738e-7, mean chord 1.734e-7 (0.3 percent); Z/X 0.01: I-HICUP 1.680e-7, mean chord 1.621e-7 (3.6 percent); Z/X 0.1: I-HICUP 1.150e-7, mean chord 8.627e-8 (33.4 percent); Z/X 0.3: I-HICUP 6.255e-8, mean chord 4.204e-8 (48.8 percent); Z/X 1: I-HICUP 2.115e-8, mean chord 1.069e-8 (98.0 percent). PASS Z/X = 0.001: I-HICUP and the mean chord effective flux agree within 1 percent (both are the cosine law there): Z/X 0.001: I-HICUP 1.738e-7, mean chord 1.734e-7 (0.3 percent) The mean chord flux falls short as the cell thickens because the chords longer than the mean, up to the diagonal, let the far more numerous lower LET ions upset; the full distribution is what CREME96's integral RPP integrates, and I-HICUP is its closed form. PASS Z/X = 0.001: I-HICUP equals the alpha law tool's cosine law rate (the DIAL_E checked integrator) within 1 percent: I-HICUP 1.7334e-7, cosine law 1.7400e-7 upsets per bit day Equation (22) for RK1-05 in GCR solar minimum, percent change of the rate for plus 10 / minus 10 percent in each element of (A, alpha, h, Sf, E0, b, W): A 36.2 / -30.5; alpha 0.3 / 0.3; h 24.2 / -20.9; Sf 5.5 / -5.2; E0 -44.4 / 71.2; b -17.9 / 22.9; W -16.4 / 21.5. Figure 11 and 12 report alpha as the least sensitive and E0 as the most sensitive element for this device. PASS sensitivity ordering as in Figures 11 and 12: alpha smallest, E0 largest
PASS fit of 28 synthetic points (1e5 bits, 1e6 per cm2, 5 percent systematic) recovers h, L0, WL, b within 15 percent and Sf within 0.3 um, chi2/dof 0.93: A 4.035e-7 vs 3.900e-7; alpha (fixed 1); h 2.89 vs 2.90; Sf 0.49 vs 0.50; L0 37.86 vs 38.00; WL 44.59 vs 43.40; b 2.08 vs 2.04 PASS bootstrap (40 replicates) 90 percent intervals contain the truth for h, Sf, L0, WL, b: h [2.63, 3.07]; Sf [0.38, 0.59]; L0 [37.51, 38.23]; WL [42.49, 47.82]; b [1.99, 2.17] PASS four more noise seeds recover h, L0, WL, b within 15 percent and Sf within 0.3 um, and the fit is never worse than the truth: seed 7: h 3.05, Sf 0.45, L0 38.0, WL 41.3, b 2.07, chi2 28.2 (truth 32.9); seed 13: h 2.70, Sf 0.50, L0 37.3, WL 40.4, b 2.26, chi2 19.6 (truth 32.4); seed 42: h 2.96, Sf 0.43, L0 37.8, WL 44.0, b 2.11, chi2 10.6 (truth 15.9); seed 77: h 2.95, Sf 0.45, L0 37.6, WL 41.9, b 2.13, chi2 23.6 (truth 29.8)
25 points, tilts 0, 45, 60 at two rotations. Alpha law (A, B free): chi2 174.8 with 6 parameters, A = 2.27, B = 0.26, GCR solar minimum rate 2.786e-11 per bit day; cosine law chi2 996.2, rate 7.942e-11. HICUP, Sf held at 0 and the aspect ratio free: chi2 173.2 with 6 parameters, RPP 366 by 36 by 83 nm (alpha 10.0, h / w 2.26), L0 1.36, WL 38.4, b 2.00, I-HICUP rate 2.680e-11 per bit day. With a square cell forced (alpha = 1): chi2 559.4, the same failure as the axisymmetric alpha law, so the azimuthal asymmetry both models find is in the data. Point by point the two fits pull the same way on 24 of 25 points (pulls within 0.7 of each other), so the residual structure is the Weibull toe at normal incidence, not the angular law; the two rates differ by 4 percent.
PASS rate from the 120 point effective flux table equals I-HICUP for the same box (three Z/X, three Weibulls) within 5 percent, the table's own interpolation error for a cube at a high threshold: Z/X 1 L0 2: -0.5%; Z/X 1 L0 10: -1.2%; Z/X 1 L0 30: -4.6%; Z/X 0.158 L0 2: -0.4%; Z/X 0.158 L0 10: -0.5%; Z/X 0.158 L0 30: -0.6%; Z/X 0.01 L0 2: -0.6%; Z/X 0.01 L0 10: 3.6%; Z/X 0.01 L0 30: 4.4% PASS far below the lowest LET every ion upsets and the effective flux is the flux through the mean projected area, (1 + 2 Z/X) / 2 of the omnidirectional flux: cube 1.499 (1.5), thin 0.510 (0.51)
All checks passed.
The transcription of the dual-node model of JPL Publication 11-6 (dualnode/dualnode.js): the lens area, the reductions, the invariant integral, the report's printed directional average and rates, a refit of its Fig. 10 data against its Table 2, and parameter recovery from synthetic data. Regenerated by dualnode/test_dualnode.js.
Generated by test_dualnode.js on 2026-09-14. 33 checks pass, 0 fail.
PASS lens area matches a 200k point Monte Carlo over 12 random circle pairs (within four sigma of the counting noise): worst difference 0.39 of the tolerance PASS D = 0 gives pi times the smaller radius squared PASS D at the sum of the radii gives zero, beyond it zero PASS D at the difference of the radii gives the smaller circle PASS two unit circles one radius apart share 2 acos(1/2) - sqrt(3)/2 = 1.22837: area 1.22837
PASS T = 0, h = 0: the dual-node cross section is the smaller single node alpha law cross section: worst relative difference 2.22e-16 PASS h = 0 at normal incidence: D = sqrt(A / B) T: 1.1337 um PASS equation (1): the alpha law integral of sigma / L^2 is the same in every direction: worst deviation 0.04 percent over four directions PASS the dual-node model violates it (Table 2 device: 75 degree tilt along the node line): ratios to normal incidence 4.15 along and 1.00 across the node line
PASS sigmaAvg reproduces the 41 printed values from LET 1 to 100 (the report prints three figures): worst relative difference 0.34 percent at LET 2.51 PASS and the 20 values below LET 1, where the average falls to 1e-22 cm2: worst relative difference 0.43 percent samples (LET: this code vs the report): 0.10: 9.902e-23 vs 9.90e-23; 0.32: 4.184e-16 vs 4.18e-16; 1.00: 5.981e-13 vs 5.98e-13; 3.16: 1.859e-11 vs 1.86e-11; 10.00: 1.467e-10 vs 1.47e-10; 31.62: 7.735e-10 vs 7.74e-10; 100.00: 1.523e-9 vs 1.52e-9 PASS the Fig. 21 spreadsheet recipe on the printed columns gives 1.49e-9 per bit day: 1.491e-9 PASS the same recipe with this code's average: 1.492e-9 PASS the site's GCR solar minimum spectrum matches the report's flux column, LET 0.3 to 25: worst difference 1.2 percent PASS the trapezoid on the report's 61 points and DN.rate on the site's spectrum (400 point grid) agree within 3 percent: 1.322e-9 and 1.332e-9 per bit day; the report's upper edge sum, 1.49e-9, is 12 percent above the converged value PASS DN.rate on the site's spectrum is within 15 percent of the report's 1.49e-9 (the difference is the bin edge rule): 1.332e-9 per bit day
PASS Table 4 (all-1's, T = 1.079): the report's spreadsheet gives 2.05e-10 per bit day; its bin edge rule on the site's spectrum reproduces that within 6 percent: edge rule 2.044e-10, converged 1.791e-10 PASS Table 6 (all-0's, T = 4.010): 5.33e-11 per bit day by the same rule: edge rule 5.314e-11, converged 4.618e-11 PASS Table 4 hand off: a Weibull fitted to the directional average over LET 1 to 30 reproduces it within 15 percent on 1.5 to 25, and the cube side is sqrt(limiting cross section): limiting 0.218 um2, cube 0.467 um (the report: 0.310 um2, 0.557 um from a fit over 1 to 100), worst deviation 13.5 percent PASS Table 6 hand off: the same on the all-0's curve: limiting 0.033 um2, cube 0.181 um (the report: 0.80 um2, 0.894 um), worst deviation 5.1 percent (the report's CREME96 runs with its Weibulls gave 1.63e-10 and 4.2e-11 against its spreadsheet sums of 2.05e-10 and 5.33e-11; the cube is as isotropic as an RPP gets and the Weibull window is a choice, so the hand off is a 20 percent step)
T sweep (um: chi2): 1: 142734494.9, 2: 875861.4, 3: 393.5, 4: 416.7, 6: 2107.6; 43 points with 30 percent errors best: T 3, A 3.602, B 0.020, S1 10.824 um2, S2 10.824 um2, h 0.033 um, phioff -10.0 deg, chi2 393.5 for 37 dof PASS the fitted model's directional average at LET 58.5 is within a factor of 2.5 of the brute force integration of the same data (6.27e-8 cm2), the report's weighting: 2.655e-8 (the report's alignment measure estimate was 3.84e-8, bounds 2.33e-8 to 6.33e-8) PASS the fit keeps the signature: the 80 degree, 180 degree spike (data 1.80e-7) within a factor of 2, and nothing at 65 degrees across the node line (data 1.22e-12): 1.46e-7 and 0.00e+0 PASS the counting error objective with a free floor gives an average within a factor of 3 of the report's weighting: 1.165e-8, floor 2.65e-12 cm2 PASS and zero at normal incidence: 0.00e+0
106 rows, 1/L^3 weighting (Section VIII); fit A 0.2372 B 0.6911 P 0.6838 L01 0.00 L02 49.59 S01 3.081 S02 1.299 um2, h -0.0106, phioff -3.1343; Table 2: A 0.2234 B 0.6953 P 0.7075 L01 0 L02 43.85 S01 2.914 S02 1.215, h 0.0066, phioff 0.0169 PASS the refit from random starts reaches the report's minimum: loss within 2 percent of the loss at the Table 2 parameters: 5.818e-6 vs 5.772e-6 PASS A, B and P within 10 percent of Table 2 PASS the two parameter sets predict the same directional cross sections where the data are (rms log ratio below 0.35: the same loss along the weakly determined direction the report describes): 0.258 over 52 directions PASS and the GCR rate from the refit is within 20 percent of the report's 1.49e-9 (which carries the bin edge rule): 1.237e-9
68 synthetic points; fitted A 0.327 B 0.766 S1 0.326 S2 1.939 L01 1.09 L02 2.66 P 1.26 h -0.007 phioff -170.2 deg, chi2/dof 0.65 PASS A and B recovered within 15 percent PASS the rotation offset recovered within 3 degrees (modulo 180): -180.2 degrees PASS chi2 per degree of freedom near 1: 0.65 PASS the GCR rate from the fit within 20 percent of the truth's: 4.274e-9 vs 4.352e-9
PASS the 45 x 180 grid of the listing agrees with a 90 x 360 grid to 1 percent: 0.00 percent the 15 x 60 grid used for bootstrap replicates differs by 0.0 percent at LET 10
